2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/export.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
94 #include <asm/tlbflush.h>
95 #include <asm/uaccess.h>
96 #include <linux/random.h>
101 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
102 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
104 static struct kmem_cache *policy_cache;
105 static struct kmem_cache *sn_cache;
107 /* Highest zone. An specific allocation for a zone below that is not
109 enum zone_type policy_zone = 0;
112 * run-time system-wide default policy => local allocation
114 static struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
120 static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
123 * If read-side task has no lock to protect task->mempolicy, write-side
124 * task will rebind the task->mempolicy by two step. The first step is
125 * setting all the newly nodes, and the second step is cleaning all the
126 * disallowed nodes. In this way, we can avoid finding no node to alloc
128 * If we have a lock to protect task->mempolicy in read-side, we do
132 * MPOL_REBIND_ONCE - do rebind work at once
133 * MPOL_REBIND_STEP1 - set all the newly nodes
134 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
136 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
137 enum mpol_rebind_step step);
138 } mpol_ops[MPOL_MAX];
140 /* Check that the nodemask contains at least one populated zone */
141 static int is_valid_nodemask(const nodemask_t *nodemask)
145 for_each_node_mask(nd, *nodemask) {
148 for (k = 0; k <= policy_zone; k++) {
149 z = &NODE_DATA(nd)->node_zones[k];
150 if (z->present_pages > 0)
158 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
160 return pol->flags & MPOL_MODE_FLAGS;
163 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
164 const nodemask_t *rel)
167 nodes_fold(tmp, *orig, nodes_weight(*rel));
168 nodes_onto(*ret, tmp, *rel);
171 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
173 if (nodes_empty(*nodes))
175 pol->v.nodes = *nodes;
179 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
182 pol->flags |= MPOL_F_LOCAL; /* local allocation */
183 else if (nodes_empty(*nodes))
184 return -EINVAL; /* no allowed nodes */
186 pol->v.preferred_node = first_node(*nodes);
190 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
192 if (!is_valid_nodemask(nodes))
194 pol->v.nodes = *nodes;
199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
200 * any, for the new policy. mpol_new() has already validated the nodes
201 * parameter with respect to the policy mode and flags. But, we need to
202 * handle an empty nodemask with MPOL_PREFERRED here.
204 * Must be called holding task's alloc_lock to protect task's mems_allowed
205 * and mempolicy. May also be called holding the mmap_semaphore for write.
207 static int mpol_set_nodemask(struct mempolicy *pol,
208 const nodemask_t *nodes, struct nodemask_scratch *nsc)
212 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
215 /* Check N_HIGH_MEMORY */
216 nodes_and(nsc->mask1,
217 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
220 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
221 nodes = NULL; /* explicit local allocation */
223 if (pol->flags & MPOL_F_RELATIVE_NODES)
224 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
226 nodes_and(nsc->mask2, *nodes, nsc->mask1);
228 if (mpol_store_user_nodemask(pol))
229 pol->w.user_nodemask = *nodes;
231 pol->w.cpuset_mems_allowed =
232 cpuset_current_mems_allowed;
236 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
238 ret = mpol_ops[pol->mode].create(pol, NULL);
243 * This function just creates a new policy, does some check and simple
244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
246 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
249 struct mempolicy *policy;
251 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
252 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
254 if (mode == MPOL_DEFAULT) {
255 if (nodes && !nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
257 return NULL; /* simply delete any existing policy */
262 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
263 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
264 * All other modes require a valid pointer to a non-empty nodemask.
266 if (mode == MPOL_PREFERRED) {
267 if (nodes_empty(*nodes)) {
268 if (((flags & MPOL_F_STATIC_NODES) ||
269 (flags & MPOL_F_RELATIVE_NODES)))
270 return ERR_PTR(-EINVAL);
272 } else if (nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
274 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
276 return ERR_PTR(-ENOMEM);
277 atomic_set(&policy->refcnt, 1);
279 policy->flags = flags;
284 /* Slow path of a mpol destructor. */
285 void __mpol_put(struct mempolicy *p)
287 if (!atomic_dec_and_test(&p->refcnt))
289 kmem_cache_free(policy_cache, p);
292 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
293 enum mpol_rebind_step step)
299 * MPOL_REBIND_ONCE - do rebind work at once
300 * MPOL_REBIND_STEP1 - set all the newly nodes
301 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
303 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
308 if (pol->flags & MPOL_F_STATIC_NODES)
309 nodes_and(tmp, pol->w.user_nodemask, *nodes);
310 else if (pol->flags & MPOL_F_RELATIVE_NODES)
311 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
314 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
317 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
318 nodes_remap(tmp, pol->v.nodes,
319 pol->w.cpuset_mems_allowed, *nodes);
320 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
321 } else if (step == MPOL_REBIND_STEP2) {
322 tmp = pol->w.cpuset_mems_allowed;
323 pol->w.cpuset_mems_allowed = *nodes;
328 if (nodes_empty(tmp))
331 if (step == MPOL_REBIND_STEP1)
332 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
333 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
338 if (!node_isset(current->il_next, tmp)) {
339 current->il_next = next_node(current->il_next, tmp);
340 if (current->il_next >= MAX_NUMNODES)
341 current->il_next = first_node(tmp);
342 if (current->il_next >= MAX_NUMNODES)
343 current->il_next = numa_node_id();
347 static void mpol_rebind_preferred(struct mempolicy *pol,
348 const nodemask_t *nodes,
349 enum mpol_rebind_step step)
353 if (pol->flags & MPOL_F_STATIC_NODES) {
354 int node = first_node(pol->w.user_nodemask);
356 if (node_isset(node, *nodes)) {
357 pol->v.preferred_node = node;
358 pol->flags &= ~MPOL_F_LOCAL;
360 pol->flags |= MPOL_F_LOCAL;
361 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 pol->v.preferred_node = first_node(tmp);
364 } else if (!(pol->flags & MPOL_F_LOCAL)) {
365 pol->v.preferred_node = node_remap(pol->v.preferred_node,
366 pol->w.cpuset_mems_allowed,
368 pol->w.cpuset_mems_allowed = *nodes;
373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
375 * If read-side task has no lock to protect task->mempolicy, write-side
376 * task will rebind the task->mempolicy by two step. The first step is
377 * setting all the newly nodes, and the second step is cleaning all the
378 * disallowed nodes. In this way, we can avoid finding no node to alloc
380 * If we have a lock to protect task->mempolicy in read-side, we do
384 * MPOL_REBIND_ONCE - do rebind work at once
385 * MPOL_REBIND_STEP1 - set all the newly nodes
386 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
388 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
389 enum mpol_rebind_step step)
393 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
394 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
397 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
400 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
403 if (step == MPOL_REBIND_STEP1)
404 pol->flags |= MPOL_F_REBINDING;
405 else if (step == MPOL_REBIND_STEP2)
406 pol->flags &= ~MPOL_F_REBINDING;
407 else if (step >= MPOL_REBIND_NSTEP)
410 mpol_ops[pol->mode].rebind(pol, newmask, step);
414 * Wrapper for mpol_rebind_policy() that just requires task
415 * pointer, and updates task mempolicy.
417 * Called with task's alloc_lock held.
420 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
421 enum mpol_rebind_step step)
423 mpol_rebind_policy(tsk->mempolicy, new, step);
427 * Rebind each vma in mm to new nodemask.
429 * Call holding a reference to mm. Takes mm->mmap_sem during call.
432 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
434 struct vm_area_struct *vma;
436 down_write(&mm->mmap_sem);
437 for (vma = mm->mmap; vma; vma = vma->vm_next)
438 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
439 up_write(&mm->mmap_sem);
442 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
444 .rebind = mpol_rebind_default,
446 [MPOL_INTERLEAVE] = {
447 .create = mpol_new_interleave,
448 .rebind = mpol_rebind_nodemask,
451 .create = mpol_new_preferred,
452 .rebind = mpol_rebind_preferred,
455 .create = mpol_new_bind,
456 .rebind = mpol_rebind_nodemask,
460 static void migrate_page_add(struct page *page, struct list_head *pagelist,
461 unsigned long flags);
463 /* Scan through pages checking if pages follow certain conditions. */
464 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
465 unsigned long addr, unsigned long end,
466 const nodemask_t *nodes, unsigned long flags,
473 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
478 if (!pte_present(*pte))
480 page = vm_normal_page(vma, addr, *pte);
484 * vm_normal_page() filters out zero pages, but there might
485 * still be PageReserved pages to skip, perhaps in a VDSO.
486 * And we cannot move PageKsm pages sensibly or safely yet.
488 if (PageReserved(page) || PageKsm(page))
490 nid = page_to_nid(page);
491 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
495 migrate_page_add(page, private, flags);
498 } while (pte++, addr += PAGE_SIZE, addr != end);
499 pte_unmap_unlock(orig_pte, ptl);
503 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
504 unsigned long addr, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags,
511 pmd = pmd_offset(pud, addr);
513 next = pmd_addr_end(addr, end);
514 split_huge_page_pmd(vma->vm_mm, pmd);
515 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
517 if (check_pte_range(vma, pmd, addr, next, nodes,
520 } while (pmd++, addr = next, addr != end);
524 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
525 unsigned long addr, unsigned long end,
526 const nodemask_t *nodes, unsigned long flags,
532 pud = pud_offset(pgd, addr);
534 next = pud_addr_end(addr, end);
535 if (pud_none_or_clear_bad(pud))
537 if (check_pmd_range(vma, pud, addr, next, nodes,
540 } while (pud++, addr = next, addr != end);
544 static inline int check_pgd_range(struct vm_area_struct *vma,
545 unsigned long addr, unsigned long end,
546 const nodemask_t *nodes, unsigned long flags,
552 pgd = pgd_offset(vma->vm_mm, addr);
554 next = pgd_addr_end(addr, end);
555 if (pgd_none_or_clear_bad(pgd))
557 if (check_pud_range(vma, pgd, addr, next, nodes,
560 } while (pgd++, addr = next, addr != end);
565 * Check if all pages in a range are on a set of nodes.
566 * If pagelist != NULL then isolate pages from the LRU and
567 * put them on the pagelist.
569 static struct vm_area_struct *
570 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
571 const nodemask_t *nodes, unsigned long flags, void *private)
574 struct vm_area_struct *first, *vma, *prev;
577 first = find_vma(mm, start);
579 return ERR_PTR(-EFAULT);
581 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
582 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
583 if (!vma->vm_next && vma->vm_end < end)
584 return ERR_PTR(-EFAULT);
585 if (prev && prev->vm_end < vma->vm_start)
586 return ERR_PTR(-EFAULT);
588 if (!is_vm_hugetlb_page(vma) &&
589 ((flags & MPOL_MF_STRICT) ||
590 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
591 vma_migratable(vma)))) {
592 unsigned long endvma = vma->vm_end;
596 if (vma->vm_start > start)
597 start = vma->vm_start;
598 err = check_pgd_range(vma, start, endvma, nodes,
601 first = ERR_PTR(err);
611 * Apply policy to a single VMA
612 * This must be called with the mmap_sem held for writing.
614 static int vma_replace_policy(struct vm_area_struct *vma,
615 struct mempolicy *pol)
618 struct mempolicy *old;
619 struct mempolicy *new;
621 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
622 vma->vm_start, vma->vm_end, vma->vm_pgoff,
623 vma->vm_ops, vma->vm_file,
624 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
630 if (vma->vm_ops && vma->vm_ops->set_policy) {
631 err = vma->vm_ops->set_policy(vma, new);
636 old = vma->vm_policy;
637 vma->vm_policy = new; /* protected by mmap_sem */
646 /* Step 2: apply policy to a range and do splits. */
647 static int mbind_range(struct mm_struct *mm, unsigned long start,
648 unsigned long end, struct mempolicy *new_pol)
650 struct vm_area_struct *next;
651 struct vm_area_struct *prev;
652 struct vm_area_struct *vma;
655 unsigned long vmstart;
658 vma = find_vma(mm, start);
659 if (!vma || vma->vm_start > start)
663 if (start > vma->vm_start)
666 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
668 vmstart = max(start, vma->vm_start);
669 vmend = min(end, vma->vm_end);
671 if (mpol_equal(vma_policy(vma), new_pol))
674 pgoff = vma->vm_pgoff +
675 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
676 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
677 vma->anon_vma, vma->vm_file, pgoff,
684 if (vma->vm_start != vmstart) {
685 err = split_vma(vma->vm_mm, vma, vmstart, 1);
689 if (vma->vm_end != vmend) {
690 err = split_vma(vma->vm_mm, vma, vmend, 0);
694 err = vma_replace_policy(vma, new_pol);
704 * Update task->flags PF_MEMPOLICY bit: set iff non-default
705 * mempolicy. Allows more rapid checking of this (combined perhaps
706 * with other PF_* flag bits) on memory allocation hot code paths.
708 * If called from outside this file, the task 'p' should -only- be
709 * a newly forked child not yet visible on the task list, because
710 * manipulating the task flags of a visible task is not safe.
712 * The above limitation is why this routine has the funny name
713 * mpol_fix_fork_child_flag().
715 * It is also safe to call this with a task pointer of current,
716 * which the static wrapper mpol_set_task_struct_flag() does,
717 * for use within this file.
720 void mpol_fix_fork_child_flag(struct task_struct *p)
723 p->flags |= PF_MEMPOLICY;
725 p->flags &= ~PF_MEMPOLICY;
728 static void mpol_set_task_struct_flag(void)
730 mpol_fix_fork_child_flag(current);
733 /* Set the process memory policy */
734 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
737 struct mempolicy *new, *old;
738 struct mm_struct *mm = current->mm;
739 NODEMASK_SCRATCH(scratch);
745 new = mpol_new(mode, flags, nodes);
751 * prevent changing our mempolicy while show_numa_maps()
753 * Note: do_set_mempolicy() can be called at init time
757 down_write(&mm->mmap_sem);
759 ret = mpol_set_nodemask(new, nodes, scratch);
761 task_unlock(current);
763 up_write(&mm->mmap_sem);
767 old = current->mempolicy;
768 current->mempolicy = new;
769 mpol_set_task_struct_flag();
770 if (new && new->mode == MPOL_INTERLEAVE &&
771 nodes_weight(new->v.nodes))
772 current->il_next = first_node(new->v.nodes);
773 task_unlock(current);
775 up_write(&mm->mmap_sem);
780 NODEMASK_SCRATCH_FREE(scratch);
785 * Return nodemask for policy for get_mempolicy() query
787 * Called with task's alloc_lock held
789 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
792 if (p == &default_policy)
798 case MPOL_INTERLEAVE:
802 if (!(p->flags & MPOL_F_LOCAL))
803 node_set(p->v.preferred_node, *nodes);
804 /* else return empty node mask for local allocation */
811 static int lookup_node(struct mm_struct *mm, unsigned long addr)
816 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
818 err = page_to_nid(p);
824 /* Retrieve NUMA policy */
825 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
826 unsigned long addr, unsigned long flags)
829 struct mm_struct *mm = current->mm;
830 struct vm_area_struct *vma = NULL;
831 struct mempolicy *pol = current->mempolicy;
834 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
837 if (flags & MPOL_F_MEMS_ALLOWED) {
838 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
840 *policy = 0; /* just so it's initialized */
842 *nmask = cpuset_current_mems_allowed;
843 task_unlock(current);
847 if (flags & MPOL_F_ADDR) {
849 * Do NOT fall back to task policy if the
850 * vma/shared policy at addr is NULL. We
851 * want to return MPOL_DEFAULT in this case.
853 down_read(&mm->mmap_sem);
854 vma = find_vma_intersection(mm, addr, addr+1);
856 up_read(&mm->mmap_sem);
859 if (vma->vm_ops && vma->vm_ops->get_policy)
860 pol = vma->vm_ops->get_policy(vma, addr);
862 pol = vma->vm_policy;
867 pol = &default_policy; /* indicates default behavior */
869 if (flags & MPOL_F_NODE) {
870 if (flags & MPOL_F_ADDR) {
871 err = lookup_node(mm, addr);
875 } else if (pol == current->mempolicy &&
876 pol->mode == MPOL_INTERLEAVE) {
877 *policy = current->il_next;
883 *policy = pol == &default_policy ? MPOL_DEFAULT :
886 * Internal mempolicy flags must be masked off before exposing
887 * the policy to userspace.
889 *policy |= (pol->flags & MPOL_MODE_FLAGS);
893 up_read(¤t->mm->mmap_sem);
899 if (mpol_store_user_nodemask(pol)) {
900 *nmask = pol->w.user_nodemask;
903 get_policy_nodemask(pol, nmask);
904 task_unlock(current);
911 up_read(¤t->mm->mmap_sem);
915 #ifdef CONFIG_MIGRATION
919 static void migrate_page_add(struct page *page, struct list_head *pagelist,
923 * Avoid migrating a page that is shared with others.
925 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
926 if (!isolate_lru_page(page)) {
927 list_add_tail(&page->lru, pagelist);
928 inc_zone_page_state(page, NR_ISOLATED_ANON +
929 page_is_file_cache(page));
934 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
936 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
940 * Migrate pages from one node to a target node.
941 * Returns error or the number of pages not migrated.
943 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
949 struct vm_area_struct *vma;
952 node_set(source, nmask);
954 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
955 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
959 if (!list_empty(&pagelist)) {
960 err = migrate_pages(&pagelist, new_node_page, dest,
961 false, MIGRATE_SYNC);
963 putback_lru_pages(&pagelist);
970 * Move pages between the two nodesets so as to preserve the physical
971 * layout as much as possible.
973 * Returns the number of page that could not be moved.
975 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
976 const nodemask_t *to, int flags)
982 err = migrate_prep();
986 down_read(&mm->mmap_sem);
988 err = migrate_vmas(mm, from, to, flags);
993 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
994 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
995 * bit in 'tmp', and return that <source, dest> pair for migration.
996 * The pair of nodemasks 'to' and 'from' define the map.
998 * If no pair of bits is found that way, fallback to picking some
999 * pair of 'source' and 'dest' bits that are not the same. If the
1000 * 'source' and 'dest' bits are the same, this represents a node
1001 * that will be migrating to itself, so no pages need move.
1003 * If no bits are left in 'tmp', or if all remaining bits left
1004 * in 'tmp' correspond to the same bit in 'to', return false
1005 * (nothing left to migrate).
1007 * This lets us pick a pair of nodes to migrate between, such that
1008 * if possible the dest node is not already occupied by some other
1009 * source node, minimizing the risk of overloading the memory on a
1010 * node that would happen if we migrated incoming memory to a node
1011 * before migrating outgoing memory source that same node.
1013 * A single scan of tmp is sufficient. As we go, we remember the
1014 * most recent <s, d> pair that moved (s != d). If we find a pair
1015 * that not only moved, but what's better, moved to an empty slot
1016 * (d is not set in tmp), then we break out then, with that pair.
1017 * Otherwise when we finish scanning from_tmp, we at least have the
1018 * most recent <s, d> pair that moved. If we get all the way through
1019 * the scan of tmp without finding any node that moved, much less
1020 * moved to an empty node, then there is nothing left worth migrating.
1024 while (!nodes_empty(tmp)) {
1029 for_each_node_mask(s, tmp) {
1032 * do_migrate_pages() tries to maintain the relative
1033 * node relationship of the pages established between
1034 * threads and memory areas.
1036 * However if the number of source nodes is not equal to
1037 * the number of destination nodes we can not preserve
1038 * this node relative relationship. In that case, skip
1039 * copying memory from a node that is in the destination
1042 * Example: [2,3,4] -> [3,4,5] moves everything.
1043 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1046 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1047 (node_isset(s, *to)))
1050 d = node_remap(s, *from, *to);
1054 source = s; /* Node moved. Memorize */
1057 /* dest not in remaining from nodes? */
1058 if (!node_isset(dest, tmp))
1064 node_clear(source, tmp);
1065 err = migrate_to_node(mm, source, dest, flags);
1072 up_read(&mm->mmap_sem);
1080 * Allocate a new page for page migration based on vma policy.
1081 * Start assuming that page is mapped by vma pointed to by @private.
1082 * Search forward from there, if not. N.B., this assumes that the
1083 * list of pages handed to migrate_pages()--which is how we get here--
1084 * is in virtual address order.
1086 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1088 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1089 unsigned long uninitialized_var(address);
1092 address = page_address_in_vma(page, vma);
1093 if (address != -EFAULT)
1099 * if !vma, alloc_page_vma() will use task or system default policy
1101 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1105 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1106 unsigned long flags)
1110 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1111 const nodemask_t *to, int flags)
1116 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1122 static long do_mbind(unsigned long start, unsigned long len,
1123 unsigned short mode, unsigned short mode_flags,
1124 nodemask_t *nmask, unsigned long flags)
1126 struct vm_area_struct *vma;
1127 struct mm_struct *mm = current->mm;
1128 struct mempolicy *new;
1131 LIST_HEAD(pagelist);
1133 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1134 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1136 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1139 if (start & ~PAGE_MASK)
1142 if (mode == MPOL_DEFAULT)
1143 flags &= ~MPOL_MF_STRICT;
1145 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1153 new = mpol_new(mode, mode_flags, nmask);
1155 return PTR_ERR(new);
1158 * If we are using the default policy then operation
1159 * on discontinuous address spaces is okay after all
1162 flags |= MPOL_MF_DISCONTIG_OK;
1164 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1165 start, start + len, mode, mode_flags,
1166 nmask ? nodes_addr(*nmask)[0] : -1);
1168 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1170 err = migrate_prep();
1175 NODEMASK_SCRATCH(scratch);
1177 down_write(&mm->mmap_sem);
1179 err = mpol_set_nodemask(new, nmask, scratch);
1180 task_unlock(current);
1182 up_write(&mm->mmap_sem);
1185 NODEMASK_SCRATCH_FREE(scratch);
1190 vma = check_range(mm, start, end, nmask,
1191 flags | MPOL_MF_INVERT, &pagelist);
1197 err = mbind_range(mm, start, end, new);
1199 if (!list_empty(&pagelist)) {
1200 nr_failed = migrate_pages(&pagelist, new_vma_page,
1202 false, MIGRATE_SYNC);
1204 putback_lru_pages(&pagelist);
1207 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1210 putback_lru_pages(&pagelist);
1212 up_write(&mm->mmap_sem);
1219 * User space interface with variable sized bitmaps for nodelists.
1222 /* Copy a node mask from user space. */
1223 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1224 unsigned long maxnode)
1227 unsigned long nlongs;
1228 unsigned long endmask;
1231 nodes_clear(*nodes);
1232 if (maxnode == 0 || !nmask)
1234 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1237 nlongs = BITS_TO_LONGS(maxnode);
1238 if ((maxnode % BITS_PER_LONG) == 0)
1241 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1243 /* When the user specified more nodes than supported just check
1244 if the non supported part is all zero. */
1245 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1246 if (nlongs > PAGE_SIZE/sizeof(long))
1248 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1250 if (get_user(t, nmask + k))
1252 if (k == nlongs - 1) {
1258 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1262 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1264 nodes_addr(*nodes)[nlongs-1] &= endmask;
1268 /* Copy a kernel node mask to user space */
1269 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1272 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1273 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1275 if (copy > nbytes) {
1276 if (copy > PAGE_SIZE)
1278 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1282 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1285 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1286 unsigned long, mode, unsigned long __user *, nmask,
1287 unsigned long, maxnode, unsigned, flags)
1291 unsigned short mode_flags;
1293 mode_flags = mode & MPOL_MODE_FLAGS;
1294 mode &= ~MPOL_MODE_FLAGS;
1295 if (mode >= MPOL_MAX)
1297 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1298 (mode_flags & MPOL_F_RELATIVE_NODES))
1300 err = get_nodes(&nodes, nmask, maxnode);
1303 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1306 /* Set the process memory policy */
1307 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1308 unsigned long, maxnode)
1312 unsigned short flags;
1314 flags = mode & MPOL_MODE_FLAGS;
1315 mode &= ~MPOL_MODE_FLAGS;
1316 if ((unsigned int)mode >= MPOL_MAX)
1318 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1320 err = get_nodes(&nodes, nmask, maxnode);
1323 return do_set_mempolicy(mode, flags, &nodes);
1326 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1327 const unsigned long __user *, old_nodes,
1328 const unsigned long __user *, new_nodes)
1330 const struct cred *cred = current_cred(), *tcred;
1331 struct mm_struct *mm = NULL;
1332 struct task_struct *task;
1333 nodemask_t task_nodes;
1337 NODEMASK_SCRATCH(scratch);
1342 old = &scratch->mask1;
1343 new = &scratch->mask2;
1345 err = get_nodes(old, old_nodes, maxnode);
1349 err = get_nodes(new, new_nodes, maxnode);
1353 /* Find the mm_struct */
1355 task = pid ? find_task_by_vpid(pid) : current;
1361 get_task_struct(task);
1366 * Check if this process has the right to modify the specified
1367 * process. The right exists if the process has administrative
1368 * capabilities, superuser privileges or the same
1369 * userid as the target process.
1371 tcred = __task_cred(task);
1372 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1373 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1374 !capable(CAP_SYS_NICE)) {
1381 task_nodes = cpuset_mems_allowed(task);
1382 /* Is the user allowed to access the target nodes? */
1383 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1388 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1393 err = security_task_movememory(task);
1397 mm = get_task_mm(task);
1398 put_task_struct(task);
1405 err = do_migrate_pages(mm, old, new,
1406 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1410 NODEMASK_SCRATCH_FREE(scratch);
1415 put_task_struct(task);
1421 /* Retrieve NUMA policy */
1422 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1423 unsigned long __user *, nmask, unsigned long, maxnode,
1424 unsigned long, addr, unsigned long, flags)
1427 int uninitialized_var(pval);
1430 if (nmask != NULL && maxnode < MAX_NUMNODES)
1433 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1438 if (policy && put_user(pval, policy))
1442 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1447 #ifdef CONFIG_COMPAT
1449 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1450 compat_ulong_t __user *nmask,
1451 compat_ulong_t maxnode,
1452 compat_ulong_t addr, compat_ulong_t flags)
1455 unsigned long __user *nm = NULL;
1456 unsigned long nr_bits, alloc_size;
1457 DECLARE_BITMAP(bm, MAX_NUMNODES);
1459 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1460 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1463 nm = compat_alloc_user_space(alloc_size);
1465 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1467 if (!err && nmask) {
1468 unsigned long copy_size;
1469 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1470 err = copy_from_user(bm, nm, copy_size);
1471 /* ensure entire bitmap is zeroed */
1472 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1473 err |= compat_put_bitmap(nmask, bm, nr_bits);
1479 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1480 compat_ulong_t maxnode)
1483 unsigned long __user *nm = NULL;
1484 unsigned long nr_bits, alloc_size;
1485 DECLARE_BITMAP(bm, MAX_NUMNODES);
1487 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1488 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1491 err = compat_get_bitmap(bm, nmask, nr_bits);
1492 nm = compat_alloc_user_space(alloc_size);
1493 err |= copy_to_user(nm, bm, alloc_size);
1499 return sys_set_mempolicy(mode, nm, nr_bits+1);
1502 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1503 compat_ulong_t mode, compat_ulong_t __user *nmask,
1504 compat_ulong_t maxnode, compat_ulong_t flags)
1507 unsigned long __user *nm = NULL;
1508 unsigned long nr_bits, alloc_size;
1511 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1512 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1515 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1516 nm = compat_alloc_user_space(alloc_size);
1517 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1523 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1529 * get_vma_policy(@task, @vma, @addr)
1530 * @task - task for fallback if vma policy == default
1531 * @vma - virtual memory area whose policy is sought
1532 * @addr - address in @vma for shared policy lookup
1534 * Returns effective policy for a VMA at specified address.
1535 * Falls back to @task or system default policy, as necessary.
1536 * Current or other task's task mempolicy and non-shared vma policies
1537 * are protected by the task's mmap_sem, which must be held for read by
1539 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1540 * count--added by the get_policy() vm_op, as appropriate--to protect against
1541 * freeing by another task. It is the caller's responsibility to free the
1542 * extra reference for shared policies.
1544 struct mempolicy *get_vma_policy(struct task_struct *task,
1545 struct vm_area_struct *vma, unsigned long addr)
1547 struct mempolicy *pol = task->mempolicy;
1550 if (vma->vm_ops && vma->vm_ops->get_policy) {
1551 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1555 } else if (vma->vm_policy)
1556 pol = vma->vm_policy;
1559 pol = &default_policy;
1564 * Return a nodemask representing a mempolicy for filtering nodes for
1567 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1569 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1570 if (unlikely(policy->mode == MPOL_BIND) &&
1571 gfp_zone(gfp) >= policy_zone &&
1572 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1573 return &policy->v.nodes;
1578 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1579 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1582 switch (policy->mode) {
1583 case MPOL_PREFERRED:
1584 if (!(policy->flags & MPOL_F_LOCAL))
1585 nd = policy->v.preferred_node;
1589 * Normally, MPOL_BIND allocations are node-local within the
1590 * allowed nodemask. However, if __GFP_THISNODE is set and the
1591 * current node isn't part of the mask, we use the zonelist for
1592 * the first node in the mask instead.
1594 if (unlikely(gfp & __GFP_THISNODE) &&
1595 unlikely(!node_isset(nd, policy->v.nodes)))
1596 nd = first_node(policy->v.nodes);
1601 return node_zonelist(nd, gfp);
1604 /* Do dynamic interleaving for a process */
1605 static unsigned interleave_nodes(struct mempolicy *policy)
1608 struct task_struct *me = current;
1611 next = next_node(nid, policy->v.nodes);
1612 if (next >= MAX_NUMNODES)
1613 next = first_node(policy->v.nodes);
1614 if (next < MAX_NUMNODES)
1620 * Depending on the memory policy provide a node from which to allocate the
1622 * @policy must be protected by freeing by the caller. If @policy is
1623 * the current task's mempolicy, this protection is implicit, as only the
1624 * task can change it's policy. The system default policy requires no
1627 unsigned slab_node(void)
1629 struct mempolicy *policy;
1632 return numa_node_id();
1634 policy = current->mempolicy;
1635 if (!policy || policy->flags & MPOL_F_LOCAL)
1636 return numa_node_id();
1638 switch (policy->mode) {
1639 case MPOL_PREFERRED:
1641 * handled MPOL_F_LOCAL above
1643 return policy->v.preferred_node;
1645 case MPOL_INTERLEAVE:
1646 return interleave_nodes(policy);
1650 * Follow bind policy behavior and start allocation at the
1653 struct zonelist *zonelist;
1655 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1656 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1657 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1660 return zone ? zone->node : numa_node_id();
1668 /* Do static interleaving for a VMA with known offset. */
1669 static unsigned offset_il_node(struct mempolicy *pol,
1670 struct vm_area_struct *vma, unsigned long off)
1672 unsigned nnodes = nodes_weight(pol->v.nodes);
1678 return numa_node_id();
1679 target = (unsigned int)off % nnodes;
1682 nid = next_node(nid, pol->v.nodes);
1684 } while (c <= target);
1688 /* Determine a node number for interleave */
1689 static inline unsigned interleave_nid(struct mempolicy *pol,
1690 struct vm_area_struct *vma, unsigned long addr, int shift)
1696 * for small pages, there is no difference between
1697 * shift and PAGE_SHIFT, so the bit-shift is safe.
1698 * for huge pages, since vm_pgoff is in units of small
1699 * pages, we need to shift off the always 0 bits to get
1702 BUG_ON(shift < PAGE_SHIFT);
1703 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1704 off += (addr - vma->vm_start) >> shift;
1705 return offset_il_node(pol, vma, off);
1707 return interleave_nodes(pol);
1711 * Return the bit number of a random bit set in the nodemask.
1712 * (returns -1 if nodemask is empty)
1714 int node_random(const nodemask_t *maskp)
1718 w = nodes_weight(*maskp);
1720 bit = bitmap_ord_to_pos(maskp->bits,
1721 get_random_int() % w, MAX_NUMNODES);
1725 #ifdef CONFIG_HUGETLBFS
1727 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1728 * @vma = virtual memory area whose policy is sought
1729 * @addr = address in @vma for shared policy lookup and interleave policy
1730 * @gfp_flags = for requested zone
1731 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1732 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1734 * Returns a zonelist suitable for a huge page allocation and a pointer
1735 * to the struct mempolicy for conditional unref after allocation.
1736 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1737 * @nodemask for filtering the zonelist.
1739 * Must be protected by get_mems_allowed()
1741 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1742 gfp_t gfp_flags, struct mempolicy **mpol,
1743 nodemask_t **nodemask)
1745 struct zonelist *zl;
1747 *mpol = get_vma_policy(current, vma, addr);
1748 *nodemask = NULL; /* assume !MPOL_BIND */
1750 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1751 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1752 huge_page_shift(hstate_vma(vma))), gfp_flags);
1754 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1755 if ((*mpol)->mode == MPOL_BIND)
1756 *nodemask = &(*mpol)->v.nodes;
1762 * init_nodemask_of_mempolicy
1764 * If the current task's mempolicy is "default" [NULL], return 'false'
1765 * to indicate default policy. Otherwise, extract the policy nodemask
1766 * for 'bind' or 'interleave' policy into the argument nodemask, or
1767 * initialize the argument nodemask to contain the single node for
1768 * 'preferred' or 'local' policy and return 'true' to indicate presence
1769 * of non-default mempolicy.
1771 * We don't bother with reference counting the mempolicy [mpol_get/put]
1772 * because the current task is examining it's own mempolicy and a task's
1773 * mempolicy is only ever changed by the task itself.
1775 * N.B., it is the caller's responsibility to free a returned nodemask.
1777 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1779 struct mempolicy *mempolicy;
1782 if (!(mask && current->mempolicy))
1786 mempolicy = current->mempolicy;
1787 switch (mempolicy->mode) {
1788 case MPOL_PREFERRED:
1789 if (mempolicy->flags & MPOL_F_LOCAL)
1790 nid = numa_node_id();
1792 nid = mempolicy->v.preferred_node;
1793 init_nodemask_of_node(mask, nid);
1798 case MPOL_INTERLEAVE:
1799 *mask = mempolicy->v.nodes;
1805 task_unlock(current);
1812 * mempolicy_nodemask_intersects
1814 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1815 * policy. Otherwise, check for intersection between mask and the policy
1816 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1817 * policy, always return true since it may allocate elsewhere on fallback.
1819 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1821 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1822 const nodemask_t *mask)
1824 struct mempolicy *mempolicy;
1830 mempolicy = tsk->mempolicy;
1834 switch (mempolicy->mode) {
1835 case MPOL_PREFERRED:
1837 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1838 * allocate from, they may fallback to other nodes when oom.
1839 * Thus, it's possible for tsk to have allocated memory from
1844 case MPOL_INTERLEAVE:
1845 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1855 /* Allocate a page in interleaved policy.
1856 Own path because it needs to do special accounting. */
1857 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1860 struct zonelist *zl;
1863 zl = node_zonelist(nid, gfp);
1864 page = __alloc_pages(gfp, order, zl);
1865 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1866 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1871 * alloc_pages_vma - Allocate a page for a VMA.
1874 * %GFP_USER user allocation.
1875 * %GFP_KERNEL kernel allocations,
1876 * %GFP_HIGHMEM highmem/user allocations,
1877 * %GFP_FS allocation should not call back into a file system.
1878 * %GFP_ATOMIC don't sleep.
1880 * @order:Order of the GFP allocation.
1881 * @vma: Pointer to VMA or NULL if not available.
1882 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1884 * This function allocates a page from the kernel page pool and applies
1885 * a NUMA policy associated with the VMA or the current process.
1886 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1887 * mm_struct of the VMA to prevent it from going away. Should be used for
1888 * all allocations for pages that will be mapped into
1889 * user space. Returns NULL when no page can be allocated.
1891 * Should be called with the mm_sem of the vma hold.
1894 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1895 unsigned long addr, int node)
1897 struct mempolicy *pol;
1898 struct zonelist *zl;
1900 unsigned int cpuset_mems_cookie;
1903 pol = get_vma_policy(current, vma, addr);
1904 cpuset_mems_cookie = get_mems_allowed();
1906 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1909 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1911 page = alloc_page_interleave(gfp, order, nid);
1912 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1917 zl = policy_zonelist(gfp, pol, node);
1918 if (unlikely(mpol_needs_cond_ref(pol))) {
1920 * slow path: ref counted shared policy
1922 struct page *page = __alloc_pages_nodemask(gfp, order,
1923 zl, policy_nodemask(gfp, pol));
1925 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1930 * fast path: default or task policy
1932 page = __alloc_pages_nodemask(gfp, order, zl,
1933 policy_nodemask(gfp, pol));
1934 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1940 * alloc_pages_current - Allocate pages.
1943 * %GFP_USER user allocation,
1944 * %GFP_KERNEL kernel allocation,
1945 * %GFP_HIGHMEM highmem allocation,
1946 * %GFP_FS don't call back into a file system.
1947 * %GFP_ATOMIC don't sleep.
1948 * @order: Power of two of allocation size in pages. 0 is a single page.
1950 * Allocate a page from the kernel page pool. When not in
1951 * interrupt context and apply the current process NUMA policy.
1952 * Returns NULL when no page can be allocated.
1954 * Don't call cpuset_update_task_memory_state() unless
1955 * 1) it's ok to take cpuset_sem (can WAIT), and
1956 * 2) allocating for current task (not interrupt).
1958 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1960 struct mempolicy *pol = current->mempolicy;
1962 unsigned int cpuset_mems_cookie;
1964 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1965 pol = &default_policy;
1968 cpuset_mems_cookie = get_mems_allowed();
1971 * No reference counting needed for current->mempolicy
1972 * nor system default_policy
1974 if (pol->mode == MPOL_INTERLEAVE)
1975 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1977 page = __alloc_pages_nodemask(gfp, order,
1978 policy_zonelist(gfp, pol, numa_node_id()),
1979 policy_nodemask(gfp, pol));
1981 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1986 EXPORT_SYMBOL(alloc_pages_current);
1989 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1990 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1991 * with the mems_allowed returned by cpuset_mems_allowed(). This
1992 * keeps mempolicies cpuset relative after its cpuset moves. See
1993 * further kernel/cpuset.c update_nodemask().
1995 * current's mempolicy may be rebinded by the other task(the task that changes
1996 * cpuset's mems), so we needn't do rebind work for current task.
1999 /* Slow path of a mempolicy duplicate */
2000 struct mempolicy *__mpol_dup(struct mempolicy *old)
2002 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2005 return ERR_PTR(-ENOMEM);
2007 /* task's mempolicy is protected by alloc_lock */
2008 if (old == current->mempolicy) {
2011 task_unlock(current);
2016 if (current_cpuset_is_being_rebound()) {
2017 nodemask_t mems = cpuset_mems_allowed(current);
2018 if (new->flags & MPOL_F_REBINDING)
2019 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2021 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2024 atomic_set(&new->refcnt, 1);
2029 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2030 * eliminate the * MPOL_F_* flags that require conditional ref and
2031 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2032 * after return. Use the returned value.
2034 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2035 * policy lookup, even if the policy needs/has extra ref on lookup.
2036 * shmem_readahead needs this.
2038 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2039 struct mempolicy *frompol)
2041 if (!mpol_needs_cond_ref(frompol))
2045 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2046 __mpol_put(frompol);
2050 /* Slow path of a mempolicy comparison */
2051 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2055 if (a->mode != b->mode)
2057 if (a->flags != b->flags)
2059 if (mpol_store_user_nodemask(a))
2060 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2066 case MPOL_INTERLEAVE:
2067 return !!nodes_equal(a->v.nodes, b->v.nodes);
2068 case MPOL_PREFERRED:
2069 return a->v.preferred_node == b->v.preferred_node;
2077 * Shared memory backing store policy support.
2079 * Remember policies even when nobody has shared memory mapped.
2080 * The policies are kept in Red-Black tree linked from the inode.
2081 * They are protected by the sp->lock spinlock, which should be held
2082 * for any accesses to the tree.
2085 /* lookup first element intersecting start-end */
2086 /* Caller holds sp->lock */
2087 static struct sp_node *
2088 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2090 struct rb_node *n = sp->root.rb_node;
2093 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2095 if (start >= p->end)
2097 else if (end <= p->start)
2105 struct sp_node *w = NULL;
2106 struct rb_node *prev = rb_prev(n);
2109 w = rb_entry(prev, struct sp_node, nd);
2110 if (w->end <= start)
2114 return rb_entry(n, struct sp_node, nd);
2117 /* Insert a new shared policy into the list. */
2118 /* Caller holds sp->lock */
2119 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2121 struct rb_node **p = &sp->root.rb_node;
2122 struct rb_node *parent = NULL;
2127 nd = rb_entry(parent, struct sp_node, nd);
2128 if (new->start < nd->start)
2130 else if (new->end > nd->end)
2131 p = &(*p)->rb_right;
2135 rb_link_node(&new->nd, parent, p);
2136 rb_insert_color(&new->nd, &sp->root);
2137 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2138 new->policy ? new->policy->mode : 0);
2141 /* Find shared policy intersecting idx */
2143 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2145 struct mempolicy *pol = NULL;
2148 if (!sp->root.rb_node)
2150 spin_lock(&sp->lock);
2151 sn = sp_lookup(sp, idx, idx+1);
2153 mpol_get(sn->policy);
2156 spin_unlock(&sp->lock);
2160 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2162 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2163 rb_erase(&n->nd, &sp->root);
2164 mpol_put(n->policy);
2165 kmem_cache_free(sn_cache, n);
2168 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2169 struct mempolicy *pol)
2172 struct mempolicy *newpol;
2174 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2178 newpol = mpol_dup(pol);
2179 if (IS_ERR(newpol)) {
2180 kmem_cache_free(sn_cache, n);
2183 newpol->flags |= MPOL_F_SHARED;
2192 /* Replace a policy range. */
2193 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2194 unsigned long end, struct sp_node *new)
2196 struct sp_node *n, *new2 = NULL;
2199 spin_lock(&sp->lock);
2200 n = sp_lookup(sp, start, end);
2201 /* Take care of old policies in the same range. */
2202 while (n && n->start < end) {
2203 struct rb_node *next = rb_next(&n->nd);
2204 if (n->start >= start) {
2210 /* Old policy spanning whole new range. */
2213 spin_unlock(&sp->lock);
2214 new2 = sp_alloc(end, n->end, n->policy);
2220 sp_insert(sp, new2);
2228 n = rb_entry(next, struct sp_node, nd);
2232 spin_unlock(&sp->lock);
2234 mpol_put(new2->policy);
2235 kmem_cache_free(sn_cache, new2);
2241 * mpol_shared_policy_init - initialize shared policy for inode
2242 * @sp: pointer to inode shared policy
2243 * @mpol: struct mempolicy to install
2245 * Install non-NULL @mpol in inode's shared policy rb-tree.
2246 * On entry, the current task has a reference on a non-NULL @mpol.
2247 * This must be released on exit.
2248 * This is called at get_inode() calls and we can use GFP_KERNEL.
2250 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2254 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2255 spin_lock_init(&sp->lock);
2258 struct vm_area_struct pvma;
2259 struct mempolicy *new;
2260 NODEMASK_SCRATCH(scratch);
2264 /* contextualize the tmpfs mount point mempolicy */
2265 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2267 goto free_scratch; /* no valid nodemask intersection */
2270 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2271 task_unlock(current);
2275 /* Create pseudo-vma that contains just the policy */
2276 memset(&pvma, 0, sizeof(struct vm_area_struct));
2277 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2278 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2281 mpol_put(new); /* drop initial ref */
2283 NODEMASK_SCRATCH_FREE(scratch);
2285 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2289 int mpol_set_shared_policy(struct shared_policy *info,
2290 struct vm_area_struct *vma, struct mempolicy *npol)
2293 struct sp_node *new = NULL;
2294 unsigned long sz = vma_pages(vma);
2296 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2298 sz, npol ? npol->mode : -1,
2299 npol ? npol->flags : -1,
2300 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2303 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2307 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2309 kmem_cache_free(sn_cache, new);
2313 /* Free a backing policy store on inode delete. */
2314 void mpol_free_shared_policy(struct shared_policy *p)
2317 struct rb_node *next;
2319 if (!p->root.rb_node)
2321 spin_lock(&p->lock);
2322 next = rb_first(&p->root);
2324 n = rb_entry(next, struct sp_node, nd);
2325 next = rb_next(&n->nd);
2326 rb_erase(&n->nd, &p->root);
2327 mpol_put(n->policy);
2328 kmem_cache_free(sn_cache, n);
2330 spin_unlock(&p->lock);
2333 /* assumes fs == KERNEL_DS */
2334 void __init numa_policy_init(void)
2336 nodemask_t interleave_nodes;
2337 unsigned long largest = 0;
2338 int nid, prefer = 0;
2340 policy_cache = kmem_cache_create("numa_policy",
2341 sizeof(struct mempolicy),
2342 0, SLAB_PANIC, NULL);
2344 sn_cache = kmem_cache_create("shared_policy_node",
2345 sizeof(struct sp_node),
2346 0, SLAB_PANIC, NULL);
2349 * Set interleaving policy for system init. Interleaving is only
2350 * enabled across suitably sized nodes (default is >= 16MB), or
2351 * fall back to the largest node if they're all smaller.
2353 nodes_clear(interleave_nodes);
2354 for_each_node_state(nid, N_HIGH_MEMORY) {
2355 unsigned long total_pages = node_present_pages(nid);
2357 /* Preserve the largest node */
2358 if (largest < total_pages) {
2359 largest = total_pages;
2363 /* Interleave this node? */
2364 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2365 node_set(nid, interleave_nodes);
2368 /* All too small, use the largest */
2369 if (unlikely(nodes_empty(interleave_nodes)))
2370 node_set(prefer, interleave_nodes);
2372 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2373 printk("numa_policy_init: interleaving failed\n");
2376 /* Reset policy of current process to default */
2377 void numa_default_policy(void)
2379 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2383 * Parse and format mempolicy from/to strings
2387 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2388 * Used only for mpol_parse_str() and mpol_to_str()
2390 #define MPOL_LOCAL MPOL_MAX
2391 static const char * const policy_modes[] =
2393 [MPOL_DEFAULT] = "default",
2394 [MPOL_PREFERRED] = "prefer",
2395 [MPOL_BIND] = "bind",
2396 [MPOL_INTERLEAVE] = "interleave",
2397 [MPOL_LOCAL] = "local"
2403 * mpol_parse_str - parse string to mempolicy
2404 * @str: string containing mempolicy to parse
2405 * @mpol: pointer to struct mempolicy pointer, returned on success.
2406 * @no_context: flag whether to "contextualize" the mempolicy
2409 * <mode>[=<flags>][:<nodelist>]
2411 * if @no_context is true, save the input nodemask in w.user_nodemask in
2412 * the returned mempolicy. This will be used to "clone" the mempolicy in
2413 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2414 * mount option. Note that if 'static' or 'relative' mode flags were
2415 * specified, the input nodemask will already have been saved. Saving
2416 * it again is redundant, but safe.
2418 * On success, returns 0, else 1
2420 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2422 struct mempolicy *new = NULL;
2423 unsigned short mode;
2424 unsigned short uninitialized_var(mode_flags);
2426 char *nodelist = strchr(str, ':');
2427 char *flags = strchr(str, '=');
2431 /* NUL-terminate mode or flags string */
2433 if (nodelist_parse(nodelist, nodes))
2435 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2441 *flags++ = '\0'; /* terminate mode string */
2443 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2444 if (!strcmp(str, policy_modes[mode])) {
2448 if (mode > MPOL_LOCAL)
2452 case MPOL_PREFERRED:
2454 * Insist on a nodelist of one node only
2457 char *rest = nodelist;
2458 while (isdigit(*rest))
2464 case MPOL_INTERLEAVE:
2466 * Default to online nodes with memory if no nodelist
2469 nodes = node_states[N_HIGH_MEMORY];
2473 * Don't allow a nodelist; mpol_new() checks flags
2477 mode = MPOL_PREFERRED;
2481 * Insist on a empty nodelist
2488 * Insist on a nodelist
2497 * Currently, we only support two mutually exclusive
2500 if (!strcmp(flags, "static"))
2501 mode_flags |= MPOL_F_STATIC_NODES;
2502 else if (!strcmp(flags, "relative"))
2503 mode_flags |= MPOL_F_RELATIVE_NODES;
2508 new = mpol_new(mode, mode_flags, &nodes);
2513 /* save for contextualization */
2514 new->w.user_nodemask = nodes;
2517 NODEMASK_SCRATCH(scratch);
2520 ret = mpol_set_nodemask(new, &nodes, scratch);
2521 task_unlock(current);
2524 NODEMASK_SCRATCH_FREE(scratch);
2533 /* Restore string for error message */
2542 #endif /* CONFIG_TMPFS */
2545 * mpol_to_str - format a mempolicy structure for printing
2546 * @buffer: to contain formatted mempolicy string
2547 * @maxlen: length of @buffer
2548 * @pol: pointer to mempolicy to be formatted
2549 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2551 * Convert a mempolicy into a string.
2552 * Returns the number of characters in buffer (if positive)
2553 * or an error (negative)
2555 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2560 unsigned short mode;
2561 unsigned short flags = pol ? pol->flags : 0;
2564 * Sanity check: room for longest mode, flag and some nodes
2566 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2568 if (!pol || pol == &default_policy)
2569 mode = MPOL_DEFAULT;
2578 case MPOL_PREFERRED:
2580 if (flags & MPOL_F_LOCAL)
2581 mode = MPOL_LOCAL; /* pseudo-policy */
2583 node_set(pol->v.preferred_node, nodes);
2588 case MPOL_INTERLEAVE:
2590 nodes = pol->w.user_nodemask;
2592 nodes = pol->v.nodes;
2599 l = strlen(policy_modes[mode]);
2600 if (buffer + maxlen < p + l + 1)
2603 strcpy(p, policy_modes[mode]);
2606 if (flags & MPOL_MODE_FLAGS) {
2607 if (buffer + maxlen < p + 2)
2612 * Currently, the only defined flags are mutually exclusive
2614 if (flags & MPOL_F_STATIC_NODES)
2615 p += snprintf(p, buffer + maxlen - p, "static");
2616 else if (flags & MPOL_F_RELATIVE_NODES)
2617 p += snprintf(p, buffer + maxlen - p, "relative");
2620 if (!nodes_empty(nodes)) {
2621 if (buffer + maxlen < p + 2)
2624 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);